Long-tail data distribution exists in most real-world datasets, and the safety hazard inspection dataset is no exception. During the safety inspection process in industrial production environments, infrequent occurrence of hazard entities exhibits a long-tail distribution. The direct application of conventional recommendation algorithms may lead to suboptimal recommendation quality and popularity bias. To improve the hazard inspection recommendation performance on long-tail datasets, we construct a Graph Convolutional Network(GCN)-based recommendation and incorporate node degrees into the model training process, using contrastive learning to enhance the representation of recommended entities. Experimental results demonstrate that our method achieves significant improvements for tail entities with Recall@10 of 0.54577 (+2.95% versus baselines) and NDCG@10 of 0.37301 (+5.22% versus baselines). Comparative analysis reveals that our approach effectively captures the heterogeneity between head and tail hazards, establishing state-of-the-art performance specifically on long-tail hazard recommendations through optimization of structural awareness.